Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Make instrumentation much more robust #2

Merged
merged 5 commits into from
Aug 19, 2021
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 5 additions & 1 deletion engine/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -79,10 +79,14 @@ If you're curious, this is how the `run.py` script works. Study the source code

## Bytecode Instrumentation

Helpful resource: https://towardsdatascience.com/understanding-python-bytecode-e7edaae8734d

The engine counts the number of bytecodes used by code. It does this by inserting a function call, `__increment__()`, in between every single bytecode (each function call requires 3 bytecodes, so this increases the code size by 4x).

Python bytecode post-3.6 consists of two bytes each. The first byte corresponds to the instruction, and the second byte to the argument. In case the argument needs to be bigger than one byte, additional `EXTENDED_ARG` are inserted before, containing the higher bits of the argument. At most 3 of them are allowed per instruction.

The bytecodes operate on a stack machine. That is, most operations are performed on the top x stack elements. Here is a list of all bytecode instructions: https://docs.python.org/3/library/dis.html#python-bytecode-instructions.

The bytecode instrumentation is not perfect, because many of the bytecodes are not constant time. For example, `BUILD_STRING(i)` concatenates `i` strings, which should probably cost `i` bytecodes and not 1 bytecode. But meh, seems relatively benign.
The bytecode instrumentation is not perfect, because many of the bytecodes are not constant time. For example, `BUILD_STRING(i)` concatenates `i` strings, which should probably cost `i` bytecodes and not 1 bytecode. But meh, seems relatively benign.

Note: `is_jump_target` is never set on the extended args, but always on the root instruction. However, the jump target is the extended args instruction.
13 changes: 12 additions & 1 deletion engine/malthusia/engine/__init__.py
Original file line number Diff line number Diff line change
@@ -1,2 +1,13 @@
import logging
import os

from .game import Game, BasicViewer, GameConstants
from .container import CodeContainer
from .container import CodeContainer

logger = logging.getLogger(__name__)
logger.setLevel(level=os.environ.get("LOGLEVEL", logging.WARNING))
ch = logging.StreamHandler()
formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')
ch.setLevel(level=logging.DEBUG)
ch.setFormatter(formatter)
logger.addHandler(ch)
66 changes: 45 additions & 21 deletions engine/malthusia/engine/container/instruction.py
Original file line number Diff line number Diff line change
Expand Up @@ -2,37 +2,61 @@
from types import SimpleNamespace

class Instruction(SimpleNamespace):
def __init__(self, instruction, in_dict=None):
if in_dict is not None:
super().__init__(**in_dict)
def __init__(self, instruction, original: bool):
vals = {a: b for a,b in zip(dis.Instruction._fields, instruction)}
vals["orig_offset"] = vals["offset"]
vals["original"] = original
assert((original and vals["offset"] is not None) or (not original and vals["offset"] is None))
vals["orig_jump_target_offset"] = None
super().__init__(**vals)

def calculate_orig_jump_target_offset(self):
assert(self.offset is not None)
assert(self.original)
assert(self.is_jumper())
assert(self.offset == self.orig_offset)
if self.is_rel_jumper():
self.orig_jump_target_offset = self.orig_offset + self.arg + 2
elif self.is_abs_jumper():
self.orig_jump_target_offset = self.arg
else:
super().__init__(**{a:b for a,b in zip(dis.Instruction._fields+('jump_to', 'was_there', 'extra_extended_args'), instruction + (None, True, 0))})
assert(False)

def is_jumper(self):
return self.is_abs_jumper() or self.is_rel_jumper()
return self.is_rel_jumper() or self.is_abs_jumper()

def is_rel_jumper(self):
return self.opcode in dis.hasjrel

def is_abs_jumper(self):
return self.opcode in dis.hasjabs

def is_extended_arg(self):
return self.opcode == dis.opmap["EXTENDED_ARG"]

@classmethod
def ExtendedArgs(self, value):
return Instruction(None, in_dict={
'opcode':144, 'opname':'EXTENDED_ARGS', 'arg':value,
'argval':value, 'argrepr':value, 'offset':None,
'starts_line':None, 'is_jump_target':False, 'was_there': False,
'extra_extended_args': 0,
})

def calculate_offset(self, instructions):
# Return the offset (rel or abs) to self.jump_to in instructions
target_loc = 2 * instructions.index(self.jump_to) - 2 * self.jump_to.extra_extended_args
def ExtendedArgs(self):
return Instruction(dis.Instruction(
opcode=dis.opmap["EXTENDED_ARG"],
opname='EXTENDED_ARG',
arg=1,
argval=1,
argrepr="",
offset=None,
starts_line=None,
is_jump_target=False
), original=False)

def calculate_jump_arg(self, orig_to_curr_offset):
assert(self.is_jumper())
assert(self.orig_jump_target_offset is not None)
assert(self.orig_jump_target_offset in orig_to_curr_offset)

target_offset = orig_to_curr_offset[self.orig_jump_target_offset]

if self.is_abs_jumper():
return target_loc

self_loc = 2 * instructions.index(self)

return target_loc - self_loc - 2
return target_offset
elif self.is_rel_jumper():
return target_offset - self.offset - 2
else:
assert(False)
Loading